ACN: Adversarial Co-training Network for Brain Tumor Segmentation with Missing Modalities

 
CONTINUE READING
ACN: Adversarial Co-training Network for Brain
 Tumor Segmentation with Missing Modalities

 Yixin Wang1,2 , Yang Zhang3( ) , Yang Liu1,2 , Zihao Lin4 , Jiang Tian3 , Cheng
 Zhong3 , Zhongchao Shi3 , Jianping Fan3,5 , and Zhiqiang He1,2,3( )
 1
 Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
arXiv:2106.14591v2 [eess.IV] 29 Jun 2021

 2
 University of Chinese Academy of Sciences, Beijing, China
 wangyixin19@mails.ucas.ac.cn
 3
 Lenovo Ltd., Beijing, China
 4
 Department of Electronic & Computer Engineering, Duke University, NC, USA
 5
 Department of Computer Science, University of North Charlotte, NC, USA

 Abstract. Accurate segmentation of brain tumors from magnetic res-
 onance imaging (MRI) is clinically relevant in diagnoses, prognoses and
 surgery treatment, which requires multiple modalities to provide com-
 plementary morphological and physiopathologic information. However,
 missing modality commonly occurs due to image corruption, artifacts,
 different acquisition protocols or allergies to certain contrast agents in
 clinical practice. Though existing efforts demonstrate the possibility of a
 unified model for all missing situations, most of them perform poorly
 when more than one modality is missing. In this paper, we propose
 a novel Adversarial Co-training Network (ACN) to solve this issue, in
 which a series of independent yet related models are trained dedicated
 to each missing situation with significantly better results. Specifically,
 ACN adopts a novel co-training network, which enables a coupled learn-
 ing process for both full modality and missing modality to supplement
 each other’s domain and feature representations, and more importantly,
 to recover the ‘missing’ information of absent modalities. Then, two
 unsupervised modules, i.e., entropy and knowledge adversarial learning
 modules are proposed to minimize the domain gap while enhancing pre-
 diction reliability and encouraging the alignment of latent representa-
 tions, respectively. We also adapt modality-mutual information knowl-
 edge transfer learning to ACN to retain the rich mutual information
 among modalities. Extensive experiments on BraTS2018 dataset show
 that our proposed method significantly outperforms all state-of-the-art
 methods under any missing situation.

 Keywords: Co-training· Missing modalities· Brain tumor segmentation.

 1 Introduction
 Malignant brain tumors have become an aggressive and dangerous disease that
 leads to death worldwide. Accurate segmentation of brain tumor is crucial to
 This work was done at AI Lab, Lenovo Research.
2 Y. Wang et al.

quantitative assessment of tumor progression and surgery treatment planning.
Magnetic resonance imaging (MRI) provides various tissue contrast views and
spatial resolutions for brain examination [2,22], by which tumor regions can be
manually segmented into heterogeneous subregions (i.e., GD-enhancing tumor,
peritumoral edema, and the necrotic and non-enhancing tumor core) by compar-
ing MRI modalities with different contrast levels (i.e., T1, T1ce, T2 and Flair).
These modalities are essential as they provide each other with complementary
information about brain structure and physiopathology. In recent years, deep

 (a) (b) (c) (d) (e) (f) (g) (h) (i)

Fig. 1. Example images. (a)-(d) Four modalities in brain MRI images; (e) Ground-truth
of three brain tumors: enhancing tumor (yellow), edema (green), necrotic and non-
enhancing tumor (red); (f) Our proposed ACN method; (g) KD-Net [11]; (h) HeMIS[10];
(i) U-HVED [8].

learning has been widely adopted in medical image analysis due to its promising
performance as well as the high cost of manual diagnosis. Joint learning from
multiple modalities significantly boosts the segmentation or registration accu-
racy, and thus has been successfully applied in previous works [9,23,19,12,20,21].
In clinical settings, however, MRI sequences acquired may be partially unusable
or missing due to corruption, artifacts, incorrect machine settings, allergies to
certain contrast agents or limited scan time. In such a situation,in only a subset
of the full treatment modalities is available.
 Despite efforts made for missing modalities, existing works usually learn com-
mon representations from full modality, and build a uniform model for all pos-
sible missing situations during inference [17,16,6,10,7,8]. However, they perform
much worse when more than one modality is missing, which is quite common in
clinical practice. Therefore, instead of a ‘catch-all’ general model for all missing
situations with low accuracies, it is more valuable to train a series of independent
yet related models dedicated to each missing situation with good accuracy. Hu
et al. [11] propose a ‘dedicated’ knowledge distillation method from multi-modal
to mono-modal segmentation. However, bias easily occurs to training a single
model for a single modality. Moreover, their models are unable to learn domain
knowledge from full modality and rich features from different levels. What’s
more, the teacher model may contain information irrelevant to the mono-modal
model, which wastes the network’s learning capacity.
 To address the obstacles mentioned above, we propose a novel Adversar-
ial Co-training Network (ACN) for missing modalities, which enables a coupled
learning process from both full modality and missing modality to supplement
each other’s domain and feature representations, and more importantly, to re-
cover the ‘missing’ information. Specifically, our model consists of two learning
ACN: Adversarial Co-training Network 3

paths: a multimodal path to derive rich modality information and a unimodal
path to generate modality-specific representations. Then, a co-training approach
is introduced to establish a coupled learning process between them, which is
achieved by three major components, namely, 1) an entropy adversarial learn-
ing module (EnA) to bridge the domain gap between multimodal and unimodal
path; 2) a knowledge adversarial learning module (KnA) to encourage the align-
ment of latent representations during multimodal and unimodal learning; 3) a
modality-mutual information knowledge transfer module (MMI) to recover the
most relevant knowledge for incomplete modalities from different feature levels
via variational information maximization. The proposed ACN outperforms all
state-of-the-arts on a popular brain tumor dataset. Fig. 1 highlights our supe-
rior results under the missing condition that only one modality is avaliable. To
the best of our knowledge, this is the first attempt to introduce the concept of
unsupervised domain adaptation (UDA) to missing modalities and the first ex-
ploration of transferring knowledge from full modality to missing modality from
the perspective of information theory.

2 Method
2.1 ACN: Adversarial Co-training Network
Co-training approach [5] has been applied to tasks with multiple feature views.
Each view of the data is independently trained and provides independent and
complementary information. Similarly, the fusion of all brain MRI modalities
provides complete anatomical and functional information about brain structure
and physiopathology. Models trained on incomplete modalities can still focus on
the specific tumor or structure of brain MRI and provide important complemen-
tary information. Therefore, we propose ACN (see Fig. 2) to enable a coupled
learning process to enhance the learning ability of both multimodal and uni-
modal training. It consists of two learning paths, i.e., the multimodal path and
the unimodal path, which are responsible for deriving generic features from full
modality and the most relevant features from available incomplete modalities,
respectively. Concretely, patches from full modality are concatenated together
to generate an input Xmulti ∈ RH×W ×M to the multimodal path, where M = 4
(Flair, T2, T1 and T1ce) in our task. The unimodal path receives patches from
N available incomplete modalities, denoted by Xuni ∈ RH×W ×N . The two paths
share the same U-Net architecture and are trained independently and simultane-
ously. We distill semantic knowledge from the output distributions of both paths
and design a consistency loss term Lcon to minimize their Kullback-Leibler (KL)
divergence, which is defined as:
 1 X X m sm suc
  
 1X m u u m c
 X
 u
 Lcon= (DKL (sc ksc)+DKL (sc ksc ))= sc log u + sc log m , (1)
 C c C c sc sc

where sm u
 c and sc denote the softened logits of multimodal and unimodal path.
C is the total number of classes. Such a coupled learning process not only pro-
motes the unimodal path to learn from multimodal path but also adds necessary
4 Y. Wang et al.

 Soft
 segmentation
 map
 Full
 Multimodal Path or
 Missing?
 1 2 3 m4
 EnA 
 Full modality
 Full
 MI MI MI MI or
 Co-training Missing? 
 
 KnA 
 MMI
 1 2 3 4
 Ground-truth
 
 Unimodal Path 
 Soft
 segmentation
 Missing modality map
 
 Conv-BN-ReLU Knowledge 
 Deconv-BN-ReLU Estimation
 Copy and Concat Weight increase

Fig. 2. Overview of our proposed ACN, consisting of a multimodal path, a unimodal
path and three modules: (a) an entropy adversarial learning module (EnA); (b) a knowl-
edge adversarial learning module (KnA); (c) a modality-mutual information knowledge
transfer module (MMI). In (a), given full modality and missing modality, Den is trained
to predict domain label based on entropy map {Imulti , Iuni }. In (b), Dkn is trained to
discriminate high-level features {Rmulti , Runi } to encourage soft-alignment of knowl-
 n oK
edge. In (c), given K pairs of representations m(k) , u(k) from both paths’
 k=1
encoder layers, ‘missing’ knowledge from multiple levels is recovered via variational
information maximization.

guidance and regularization to multimodal path to further replenish the learned
generic features.
 This consistency exploits the correlations between full modality and miss-
ing modality in the output level. However, the two paths learn from different
modalities and domains. It is significant to explicitly enhance the alignment of
different domains and high-level knowledge. Therefore, we propose two adver-
sarial learning modules to minimize the distribution distance between the two
paths and utilize knowledge transfer to retain rich mutual information from a
perspective of information theory.

2.2 Entropy Adversarial Learning

The prediction of multimodal path is not necessarily more accurate than uni-
modal path. However, models trained on multiple modalities tend to be more
reliable than those on incomplete modalities. Therefore, we adopt the principle
of entropy maps as a confidence measurement from unsupervised domain adap-
tation (UDA) tasks [18,15] and design an entropy adversarial module to match
the distributions between two paths.
 In detail, the multimodal and unimodal path receive Xmulti and Xuni sepa-
rately, along
 (h,w,c)  with their corresponding pixel-level C-class ground-truth Y (h,w) =
 Y c
 . The multimodal path takes Xmulti as input and generates the soft
ACN: Adversarial Co-training Network 5
 h i
 (h,w,c)
segmentation map Pmulti at the output level, which represents the discrete
 c
distribution over C classes. It is observed that predictions from unimodal path
tend to be under-confident with high-entropy. Conversely, predictions from mul-
timodal path are usually over-confident with low-entropy. Therefore, we adopt a
unified adversarial training strategy to minimize the entropy of unimodal learn-
ing by encouraging its entropy distribution to be more similar to the multimodal
one. Concretely, the entropy map Imulti of multimodal path is defined in the same
way as weighted self-information maps [18], calculated by:
 (h,w) (h,w,c) (h,w,c)
 X
 Imulti = −Pmulti · log Pmulti . (2)
 c

Similarly, receiving missing modality patches Xuni , the unimodal pathh can be i
 (h,w,c)
seen as a generator Gen , which generates the soft segmentation map Puni
 c
and entropy map Iuni . These pixel-level vectors can be seen as the disentangle-
ment of the Shannon Entropy, which reveals the prediction confidence. We then
introduce a fully-convolutional network as a discriminator Den . The generator
Gen tries to generate an unimodal entropy map Iuni and fools Den , while the
discriminator Den aims to distinguish Iuni from the multimodal entropy map
Imulti . Accordingly, the optimization of Gen and Den is achieved by the follow-
ing objective function:
      
 (h,w) (h,w)
 Ladv
 P
 en (Xmulti , Xuni ) = log 1 − Den Iuni + log Den Imulti . (3)
 h,w

2.3 Knowledge Adversarial Learning
Considering that the high-level representations contain richer information, we
also need to encourage features distribution alignment in latent space. Simply
minimizing the KL divergence between the features of two paths’ bottlenecks
{Rmulti , Runi } may easily disturb the underlying learning of unimodal path in
the deep layers. Therefore, we encourage the high-level representations of both
paths to be aligned using a knowledge adversarial module. Similar to EnA, the
generator Gkn tries to generate high-level features to mislead another discrimi-
nator Dkn . The objective function of this process is formulated as:

 Ladv
 kn (Xmulti , Xuni ) = log (1 − Dkn (Runi )) + log (Dkn (Rmulti )) . (4)

KnA serves as a soft-alignment to encourage unimodal path to learn abundant
and ‘missing’ knowledge from full modality.

2.4 Modality-mutual Information Knowledge Transfer Learning
Multimodal path may contain information irrelevant to the task, which requires
superfluous effort for alignment by unimodal path. To address this issue, we in-
troduce the modality-mutual information knowledge transfer learning to retain
6 Y. Wang et al.

high mutual information between two paths. Specifically, K pairs of representa-
   K
tions m(k) , u(k) k=1 can be obtained from K encoder layers of the multimodal
and unimodal path separately. Given the entropy H(m) and conditional entropy
H(m | u), the mutual information M I between each pair (m, u) can be defined
by: M I(m, u) = H(m) − H(m | u), which measures a reduction in uncertainty
in the knowledge of the multimodal learning encoded in its layers when the uni-
modal knowledge is known. Following [1] to measure the exact values, we use
variational information maximization [4] for each M I(m, u):
M I(m, u) = H(m)−H(m | u) = H(m)+Em,u∼p(m,u) [log p(m | u)]
 = H(m)+Eu∼p(u) [DKL (p(m | u)kq(m | u))]+Em,u∼p(m,u) [log q (m | u)]
 ≥ H(m)+Em,u∼p(m,u) [log q(m | u)].
 (5)
The distribution p(m | u) is approximated by a variational distribution q(m | u).
Accordingly, the optimization of hierarchical mutual information can be formu-
lated with the following loss function LMI :
 K
 X   X K h  i
 LMI = − γk MI m(k),u(k) = − γk Em(k),u(k) ∼p(m(k),u(k)) log q m(k)| u(k) , (6)
 k=1 k=1

where γk increases with k, indicating that higher layers contain more semantic
information which should be assigned with larger weights for guidance. For spe-
cific implementation of our co-training network, the variation distribution can
be realized by:
 H X
 C X W 2
 X (mc,h,w −µc,h,w (u))
 − log q(m | u) = log σc + + Z, (7)
 c=1 h=1w=1
 2σc2

where µ(·) and σ denote the heteroscedastic mean and homoscedastic variance
function of a Gaussian distribution. W and H denote width and height, C is the
channel numbers of the corresponding layers and Z is a constant value.

2.5 Overall Loss and Training Procedure
The overall loss funtion L is formulated by:
 L =λmulti Ldice dice adv adv
 multi + λuni Luni + ω(t)Lcon + λ0 Len + λ1 Lkn + λ2 LM I . (8)
where Ldice dice
 multi and Luni are commonly used segmentation Dice loss in medical
 2
tasks for multimodal and unimodal path respectively. ω(t) = 0.1 ∗ e(−5(1−S/L) )
is a time-dependent Gaussian weighting function, where S and L represent the
current training step and ramp-up length separately. λmulti , λuni , λ0 , λ1 and λ2
are trade-off parameters, which are set as 0.2, 0.8, 0.001, 0.0002, and 0.5 in our
model. They are chosen according to the performance under the circumstance
that only T1ce modality is available since it’s the major modality for tumor
diagnosis. The ultimate goal of the overall co-training procedure is the following
optimization: min max L.
 Gen ,Gkn Den ,Dkn
ACN: Adversarial Co-training Network 7

3 Experimental Results

3.1 Experimental Setup

Dataset BraTS2018 training dataset [13,2,3] consists of 285 multi-contrast MRI
scans with four modalities: a) native (T1), b) post-contrast T1-weighted (T1ce),
c) T2-weighted (T2), and d) T2 Fluid Attenuated Inversion Recovery (Flair)
volumes. Each of these modalities captures different properties of brain tumor
subregions: GD-enhancing tumor (ET), the peritumoral edema (ED), and the
necrotic and non-enhancing tumor core (NCR/NET). These subregions are com-
bined into three nested subregions: whole tumor (WT), tumor core (TC) and
enhancing tumor (ET). All the volumes have been co-registered to the same
anatomical template and interpolated to the same resolution.

Implementation and Evaluation Experiments are implemented in Pytorch
and performed on NVIDIA Tesla V100 with 32GB Ram. For fair comparison,
we follow the experimental settings of the winner of BraTS2018 [14]. Both paths
share the same U-Net backbone as [14] and the discriminators Den and Dkn are
two fully-convolutional networks. The input patch size is set as 160 × 192 × 128
and batch size as 1. The adam optimizer is applied, and the initial learning
rate is set as 1e − 4 and progressively decreases according to a poly policy
 0.9
(1 − epoch / epoch max ) , where epoch max is the total number of epochs
(300). We randomly split the dataset into training and validation sets by a ratio
of 2:1. The segmentation performance is evaluated on each nested subregion of
brain tumors using ‘Dice similarity coefficient (DSC)’ and ‘Hausdorff distance
(HD95)’. A higher DSC and a lower HD95 indicate a better segmentation per-
formance. Codes will be available at https://github.com/Wangyixinxin/ACN.

3.2 Results and Analysis

Comparisons with State-of-the-art Methods We compare the proposed
ACN with three state-of-the-art methods: two ‘catch-all’ models HeMIS [10] and
U-HVED [8] and a ‘dedicated’ model KD-Net [11]. For fair comparison, U-Net is
applied as the benchmark to all methods. Table 1 shows that ACN significantly
outperforms both HeMIS and U-HVED for all the 15 possible combinations of
missing modality, especially when only one or two modalities are available. It
is noted that missing T1ce leads to a severe decreasing on both ET and TC
while missing Flair causes a significant drop on WT. Both HeMIS and U-HVED
perform poorly under the above circumstances, with unacceptable DSC scores
of only about 10% − 20% on the ET and TC. In contrast, our method achieves
promising DSC scores of above 40% in such hard situations, which is much
more valuable in clinical practice. Table 2 shows the DSC comparison with the
‘dedicated’ KD-Net. Since KD-Net aims to deal with mono-modal segmentation,
we set the comparison under the condition that only one modality is available.
8 Y. Wang et al.

Table 1. Comparison with state-of-the-art ‘catch-all’ methods (DSC %) on three
nested subregions (ET, TC and WT). Modalities present are denoted by •, the missing
ones by ◦.

 ET TC WT
 Modalities U-HeMIS U-HVED ACN U-HeMIS U-HVED ACN U-HeMIS U-HVED ACN
Flair T1 T1ce T2 DSC HD95 DSC HD95 DSC HD95 DSC HD95 DSC HD95 DSC HD95 DSC HD95 DSC HD95 DSC HD95
 ◦ ◦ ◦ • 25.63 14.42 22.82 14.28 42.98 10.66 57.20 17.88 54.67 15.38 67.94 10.07 80.96 12.53 79.83 14.63 85.55 7.24
 ◦ ◦ • ◦ 62.02 22.87 57.64 31.90 78.07 3.57 65.29 28.21 59.59 38.01 84.18 5.04 61.53 28.23 53.62 34.14 80.52 8.42
 ◦ • ◦ ◦ 10.16 26.06 8.60 28.80 41.52 10.68 37.39 30.70 33.90 32.29 71.18 10.46 57.62 27.40 49.51 31.87 79.34 10.22
 • ◦ ◦ ◦ 11.78 25.85 23.80 14.39 42.77 11.44 26.06 30.69 57.90 15.16 67.72 11.75 52.48 28.21 84.39 12.40 87.30 7.81
 ◦ ◦ • • 67.83 9.06 67.83 10.70 75.65 4.36 76.64 11.16 73.92 14.56 84.41 6.41 82.48 10.23 81.32 12.24 86.41 7.41
 ◦ • • ◦ 66.22 14.89 61.11 26.92 75.21 3.77 72.46 16.86 67.55 29.65 84.59 5.76 68.47 19.55 64.22 28.46 80.05 9.27
 • • ◦ ◦ 10.71 25.71 27.96 15.09 43.71 11.38 41.12 25.06 61.14 14.70 71.30 11.87 64.62 20.69 85.71 12.20 87.49 8.88
 ◦ • ◦ • 32.39 13.60 24.29 14.33 47.39 9.10 60.92 15.18 56.26 14.66 73.28 8.72 82.41 11.90 81.56 11.82 85.50 7.96
 • ◦ ◦ • 30.22 13.44 32.31 12.84 45.96 10.45 57.68 15.73 62.70 12.76 71.61 10.31 82.95 11.51 87.58 7.89 87.75 6.65
 • ◦ • ◦ 66.10 15.33 68.36 9.66 77.46 4.22 71.49 19.34 75.07 11.77 83.35 5.83 68.99 19.77 85.93 11.57 88.28 7.47
 • • • ◦ 68.54 12.32 68.60 9.66 76.16 5.62 76.01 12.43 77.05 10.23 84.25 6.75 72.31 14.29 86.72 11.11 88.96 6.93
 • • ◦ • 31.07 13.94 32.34 11.79 42.09 10.81 60.32 14.22 63.14 11.58 67.86 10.69 83.43 11.58 88.07 7.88 88.35 6.14
 • ◦ • • 68.72 8.03 68.93 7.72 75.97 4.34 77.53 9.02 76.75 9.21 82.85 6.48 83.85 9.26 88.09 8.00 88.34 7.03
 ◦ • • • 69.92 7.81 67.75 10.19 76.10 5.01 78.96 8.93 75.28 10.69 84.67 5.86 83.94 9.09 82.32 11.08 86.90 6.24
 • • • • 70.24 7.43 69.03 8.37 77.06 5.09 79.48 8.95 77.71 8.63 85.18 5.94 84.74 8.99 88.46 7.80 89.22 6.71
 Average 46.10 15.38 46.76 15.11 61.21 7.37 62.57 17.62 64.84 16.62 77.62 8.13 74.05 16.22 79.16 14.87 85.92 7.62

Table 2. Comparison with state-of-the-art ‘dedicated’ method KD-Net [11] (DSC %).

 Modalities ET TC WT Average
Flair T1 T1ce T2 KD-Net ACN KD-Net ACN KD-Net ACN KD-Net ACN
 ◦ ◦ ◦ • 39.04 42.98 66.01 67.94 82.32 85.55 62.46 65.49
 ◦ ◦ • ◦ 75.32 78.07 81.89 84.18 76.79 80.52 78.00 80.92
 ◦ • ◦ ◦ 39.87 41.52 70.02 71.18 77.28 79.34 62.39 64.01
 • ◦ ◦ ◦ 40.99 42.77 65.97 67.72 85.14 87.30 64.03 65.93

(a) T1ce+ GT (b) Full modality (c) Co-training (d) ACN (a) T1ce+ GT (b) Full modality (c) Co-training (d) ACN (a) T1ce+ GT (b) Full modality (c) Co-training (d) ACN

Fig. 3. Qualitative results on BraTS2018 dataset. Column (a) shows one of the input
modalities (T1ce) and the corresponding segmentation ground-truth. Column (b)(c)(d)
show the prediction entropy maps from multimodal path (full modality), unimodal path
(only T1ce modality) without ACN modules and unimodal path with ACN modules,
along with their segmentation results.
ACN: Adversarial Co-training Network 9

Ablation Study In this section, we investigate the effectiveness of each pro-
posed module, i.e., EnA, KnA and MMI. We choose a major modality T1ce as
the single available modality. First we build a co-training baseline network which
only uses consistency loss Lcon . Then we add the three modules gradually. It is
observed in Table 3 that the unimodal path for missing modality is improved
simultaneously through adding these modules, which proves the mutual benefit
of proposed co-training network and the cooperative work of the three modules.
Fig. 3 further verifies the superiority and rationality of the proposed modules.
The entropy maps indicate the prediction of missing modality is less reliable
than full modality, while our proposed ACN produces significantly better seg-
mentation for missing modality with reduced uncertainty.

Table 3. Effectiveness of each module on unimodal (T1ce modality) paths (DSC %).

 Lcon EnA KnA MMI ET TC WT Average
 √
 74.57 82.08 77.72 78.12
 √ √
 76.40 83.78 79.78 79.98
 √ √ √
 T1ce modality √ 77.14 83.85 78.82 79.94
 √ √
 76.12 84.24 79.89 80.09
 √ √ √ √
 78.07 84.18 80.52 80.92

4 Conclusion

In this work, we propose a novel Adversarial Co-training Network to address the
problem of missing modalities in brain tumor segmentation. More importantly,
we present two unsupervised adversarial learning modules to align domain and
feature distributions between full modality and missing modality. We also in-
troduce a modality-mutual information module to recover ‘missing’ knowledge
via knowledge transfer. Our model outperforms all existing methods on the mul-
timodal BraTS2018 dataset in all missing situations by a considerable margin.
Since our ‘One stop shop’ method needs to train ’dedicated’ models for each
missing situation, it may bring training cost, but it brings large improvement
during inference without extra cost, which is of great value to clinical application
and can be also generalized to other incomplete data domains.

References
 1. Ahn, S., Hu, S.X., Damianou, A., Lawrence, N.D., Dai, Z.: Variational information
 distillation for knowledge transfer. In: Proceedings of the IEEE/CVF Conference
 on Computer Vision and Pattern Recognition (CVPR) (June 2019)
 2. Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J., Freymann,
 J., Farahani, K., Davatzikos, C.: Advancing the cancer genome atlas glioma mri
10 Y. Wang et al.

 collections with expert segmentation labels and radiomic features. Scientific Data
 4 (09 2017). https://doi.org/10.1038/sdata.2017.117
 3. Bakas, S., Reyes, M., Jakab, A., Bauer, S., Rempfler, M., Crimi, A., et al.: Identify-
 ing the best machine learning algorithms for brain tumor segmentation, progression
 assessment, and overall survival prediction in the brats challenge (2018)
 4. Barber, D., Agakov, F.: The im algorithm: A variational approach to information
 maximization. (01 2003)
 5. Blum, A., Mitchell, T.: Combining labeled and unlabeled data with co-training.
 Proceedings of the Annual ACM Conference on Computational Learning Theory
 (10 2000). https://doi.org/10.1145/279943.279962
 6. Chartsias, A., Joyce, T., Giuffrida, M.V., Tsaftaris, S.A.: Multimodal mr synthesis
 via modality-invariant latent representation. IEEE Transactions on Medical Imag-
 ing 37(3), 803–814 (2018). https://doi.org/10.1109/TMI.2017.2764326
 7. Chen, C., Dou, Q., Jin, Y., Chen, H., Qin, J., Heng, P.A.: Robust multimodal
 brain tumor segmentation via feature disentanglement and gated fusion. In: Shen,
 D., Liu, T., Peters, T.M., Staib, L.H., Essert, C., Zhou, S., Yap, P.T., Khan, A.
 (eds.) Medical Image Computing and Computer Assisted Intervention – MICCAI
 2019. pp. 447–456. Springer International Publishing, Cham (2019)
 8. Dorent, R., Joutard, S., Modat, M., Ourselin, S., Vercauteren, T.: Hetero-modal
 variational encoder-decoder for joint modality completion and segmentation. In:
 Shen, D., Liu, T., Peters, T.M., Staib, L.H., Essert, C., Zhou, S., Yap, P.T., Khan,
 A. (eds.) Medical Image Computing and Computer Assisted Intervention – MIC-
 CAI 2019. pp. 74–82. Springer International Publishing, Cham (2019)
 9. Havaei, M., Davy, A., Warde-Farley, D., Biard, A., Courville, A., Bengio, Y., Pal,
 C., Jodoin, P.M., Larochelle, H.: Brain tumor segmentation with deep neural net-
 works. Medical Image Analysis 35, 18 – 31 (2017), http://www.sciencedirect.
 com/science/article/pii/S1361841516300330
10. Havaei, M., Guizard, N., Chapados, N., Bengio, Y.: Hemis: Hetero-modal image
 segmentation. In: Ourselin, S., Joskowicz, L., Sabuncu, M.R., Unal, G., Wells, W.
 (eds.) Medical Image Computing and Computer-Assisted Intervention – MICCAI
 2016. pp. 469–477. Springer International Publishing, Cham (2016)
11. Hu, M., Maillard, M., Zhang, Y., Ciceri, T., Barbera, G.L., Bloch, I., Gori, P.:
 Knowledge distillation from multi-modal to mono-modal segmentation networks.
 In: MICCAI (2020)
12. Kamnitsas, K., Ledig, C., Newcombe, V., Simpson, J., Kane, A., Menon, D.,
 Rueckert, D., Glocker, B.: Efficient multi-scale 3d cnn with fully connected crf
 for accurate brain lesion segmentation. Medical Image Analysis 36 (03 2016).
 https://doi.org/10.1016/j.media.2016.10.004
13. Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J.,
 et al.: The multimodal brain tumor image segmentation benchmark (brats). IEEE
 Transactions on Medical Imaging 34(10), 1993–2024 (2015)
14. Myronenko, A.: 3d mri brain tumor segmentation using autoencoder regularization
 (2018)
15. Pan, F., Shin, I., Rameau, F., Lee, S., Kweon, I.S.: Unsupervised intra-domain
 adaptation for semantic segmentation through self-supervision. In: Proceedings of
 the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
 (June 2020)
16. Sharma, A., Hamarneh, G.: Missing mri pulse sequence synthesis using multi-
 modal generative adversarial network. IEEE transactions on medical imag-
 ing 39(4), 1170—1183 (April 2020). https://doi.org/10.1109/tmi.2019.2945521,
 https://doi.org/10.1109/TMI.2019.2945521
ACN: Adversarial Co-training Network 11

17. van Tulder, G., de Bruijne, M.: Why does synthesized data improve multi-sequence
 classification? In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A. (eds.) Medical
 Image Computing and Computer-Assisted Intervention – MICCAI 2015. pp. 531–
 538. Springer International Publishing, Cham (2015)
18. Vu, T.H., Jain, H., Bucher, M., Cord, M., Perez, P.: Advent: Adversarial entropy
 minimization for domain adaptation in semantic segmentation. In: Proceedings of
 the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
 (June 2019)
19. Wang, Y., Zhang, Y., Hou, F., Liu, Y., Tian, J., Zhong, C., Zhang, Y., He, Z.:
 Modality-pairing learning for brain tumor segmentation. In: Crimi, A., Bakas, S.
 (eds.) Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries.
 pp. 230–240. Springer International Publishing, Cham (2021)
20. Xu, Z., Luo, J., Yan, J., Pulya, R., Li, X., Wells, W., Jagadeesan, J.: Adversar-
 ial uni- and multi-modal stream networks for multimodal image registration. In:
 Martel, A.L., Abolmaesumi, P., Stoyanov, D., Mateus, D., Zuluaga, M.A., Zhou,
 S.K., Racoceanu, D., Joskowicz, L. (eds.) Medical Image Computing and Com-
 puter Assisted Intervention – MICCAI 2020. pp. 222–232. Springer International
 Publishing, Cham (2020)
21. Xu, Z., Yan, J., Luo, J., Li, X., Jagadeesan, J.: Unsupervised multimodal image
 registration with adaptative gradient guidance. In: ICASSP 2021 - 2021 IEEE
 International Conference on Acoustics, Speech and Signal Processing (ICASSP).
 pp. 1225–1229 (2021). https://doi.org/10.1109/ICASSP39728.2021.9414320
22. Yan, J., Chen, S., Zhang, Y., Li, X.: Neural architecture search for compressed
 sensing magnetic resonance image reconstruction. Computerized Medical Imaging
 And Graphics (2020)
23. Zhou, T., Ruan, S., Canu, S.: A review: Deep learning for medi-
 cal image segmentation using multi-modality fusion. Array 3-4, 100004
 (2019). https://doi.org/https://doi.org/10.1016/j.array.2019.100004, http://www.
 sciencedirect.com/science/article/pii/S2590005619300049
You can also read